Plus plan, Codex used up 40% weekly limit on one task. Effectively unusable. You are fired! To add insult to injury, I now have to create a post of > 100 characters?
What language(s) was the code written in? Like English/Python or French/C++. Sometimes the base language can get tokenized weird, at least that is what we’ve seen in the past.
Ext │ Count │ Purpose │
├───────────────────────────────────────────────┼───────┼─────────────────────────────────────┤
│ .md │ 876 │ governance/workflow docs, specs, │
│ │ │ handoffs, archives │
├───────────────────────────────────────────────┼───────┼─────────────────────────────────────┤
│ .js │ 496 │ source + tests (Node) │
├───────────────────────────────────────────────┼───────┼─────────────────────────────────────┤
│ .mjs │ 54 │ scripts (ESM)
Ok, that’s not much code if those are line counts. What language are you talking to the model in?
No those are file counts. did some further research that (.md files) is the bottleneck. Doing more research now. Codex isn’t fired yet.
OK, let us know what you find!
Just now on my Plus account 50% of the weekly limit was used up in a single prompt in a minute.
I used the reset button to restore the limits to 100% but the next prompt reset the limits to zero again, and Codex didn’t even have time to respond.
Now I have 0% of limits and 0 resets.
Update, for anyone following: instead of ranting further, I spent the week hardening my repo’s dev workflow to cut token/context burn on both Codex and Claude Code, since that’s what triggered this post.
What changed:
- Deterministic session start/end scripts replace manual state verification with one command that prints an authoritative report.
- ~16 pre-commit validators now catch things that used to require an agent to read files and reason it out by hand (encoding compliance, half-staged file moves, handoff schema, archive-completion proof, inventory drift). Moving verification from “LLM reads and checks” to “script checks” cut the most tokens.
- Effort/model gating: cheap mechanical tasks get a cheap reasoning setting; only genuinely ambiguous work gets the expensive one.
- Audited the workflow for remaining rules that still relied on manual re-verification and converted the fixable ones into scripts/hooks this week.
None of this required switching providers — it’s workflow discipline, not model choice. Still annoyed about the 40%-in-one-task thing, but at least my own setup isn’t making it worse.
If you had to pay someone to write you 800 pages of documentation.. I don’t know what people expect.. do you even read the md files?